Understanding and enhancing visual search performance in complex scenes

描述(由申请人提供)：了解和提高复杂场景下的视觉搜索性能癌症筛查拯救生命(例如，国家肺部筛查试验研究，2011年)。每天，放射科医生都面临着困难、耗时的视觉搜索任务，比如乳房X光检查和肺癌筛查。通常，癌症的迹象，例如肺结节或乳房的微小异常，在不同的背景下很难找到，并被重叠的组织所掩盖。错过这些癌症征兆可能会导致错误的诊断，造成生死后果。因此，确定和解决在这些关键的搜索任务中提出的问题是非常重要的。拟议的研究旨在通过两种方式提高癌症筛查的搜索性能：第一，通过提高嵌入3D体积数据集中的肺结节的可见性。放射科医生通常通过滚动一叠叠的胸部CT来搜索肺结节。结节大致为球形特征，横跨CT层叠中的几个切片。据传闻，专家们报告说，肺结节在不断变化的视野中“弹出”的方式是它们存在的信号。我们提出了一种创新的方法，使用显著算法，利用这种信号来增强胸部CT，以一种将注意力引导到场景的关键区域的方式。其次，我们的目标是通过理解和利用乳房X光照片的非选择性的、类似GIST的处理来提高搜索性能。最近的研究表明，专业放射科医生可以在乳房X光照片中检测到全局信号，从而在极短的时间内暴露于刺激后，对正常和异常的乳房进行高概率分类。我们将首先培训新手，使其成为更接近医疗任务的专家。随着专业知识的发展，我们将使用脑电(EEG)研究在训练过程中可能演变的两种不同的神经关联，即P300和N2PC。在其他环境中，这些测量可以在短暂呈现的图像(P300)或(N2pc)内发出注意选择的信号。我们将采用一种新的方法，利用机器学习来实时解码大脑信号。它允许响应一系列图像而产生的神经特征被用来按照观察者对这些图像隐含的“兴趣”的顺序对它们进行排序。我们假设，这些信息可以反馈给观察者/放射科医生，作为信息的来源，例如，可能表明某个图像或区域值得更多的检查。因此，这项建议的主要目标是理解复杂视觉搜索任务中注意力的指导作用，并将这一知识应用于临床相关搜索任务的改进，如癌症筛查。

DESCRIPTION (provided by applicant): Understanding and enhancing visual search performance in complex scenes Cancer screening saves lives (e.g. National Lung Screening Trial Research, 2011). Every day, radiologists are faced with difficult, time-consuming visual search tasks like in mammography and lung cancer screening. Oftentimes signs of cancer, e.g. lung nodules or little abnormalities in a breast, are very hard to find against heterogeneous backgrounds and obscured by overlapping tissues. Missing these signs of cancer can result in wrong diagnoses with life or death consequences. It is therefore of key interest to identify and tackle the problems posed in these crucial search tasks. The proposed studies aim to improve search performance in cancer screening in two ways: First, by enhancing the visibility of lung nodules embedded in 3D volumetric datasets. Radiologists usually search for lung nodules by scrolling through stacks of chest CT. Nodules are roughly spherical features, spanning a few slices in a CT stack. Anecdotally, experts report that the way that lung nodules 'pop' in and out of the changing view is a signal to their presence. We propose an innovative approach using saliency algorithms that harness this signal to enhance chest CTs in a way that directs attention to crucial regions of a scene. Second, we aim to improve search performance by understanding and utilizing non-selective, 'gist'-like processing of mammograms. Recent work has shown that expert radiologists can detect a global signal in mammograms that allows for above-chance categorization of normal and abnormal breasts after very short exposures to the stimulus. We will first train novices to become experts in closer approximation to the medical tasks. As expertise develops, we will investigate two different neural correlates that might evolve in the course of training using electroencephalography (EEG), namely the P300 and the N2pc. In other settings, these measures can signal attentional selection either across (P300) or within (N2pc) briefly presented images. We will adapt a new method that exploits machine learning for real-time decoding of brain signals. It allows neural signatures, elicited in response to a sequence of images, to be used to rank those images in order of their implicit 'interest' to the viewer. We hypothesize that this information can be fed back to the observer/radiologist as a source of information that might, for example, suggest that an image or region deserves more scrutiny. The main goals of this proposal are therefore to understand the guidance of attention in complex visual search tasks and to apply this knowledge to improvements in clinically relevant search tasks like cancer screening.